OpenAI publishes Codex guide: how to set up workspaces, projects and task threads

On April 23, 2026, OpenAI published Working with Codex in its Academy, a basic guide for people who are just starting to use the agent for working with files and tasks. The text skips theory and covers the most practical points: how the workspace is organized, what a thread and a project are, where files are stored and how to give Codex clear instructions.

What is inside Codex

The interface centers on a sidebar that lets the user switch between threads, projects and tools. OpenAI suggests treating a thread the way you treat an ordinary chat in ChatGPT: a separate context for working on one task. A new thread can stay standalone or be placed inside a project right away. The sidebar also holds search across past tasks and projects, alongside plugins that help run repeatable processes and connect external tools.

The guide puts particular emphasis on the dialog window. OpenAI states plainly that users need neither a perfect prompt nor complex technical instructions. Codex works better when the goal, the relevant files and the definition of done are described in plain human language. That is an important signal: the product is positioned not as a narrow IDE feature for specialists, but as a working interface you can enter in everyday language and refine the task step by step.

"You don't need a perfect prompt or technical instructions."

Projects and folders

The key organizing unit in Codex is the project. Each project is linked to a specific folder on the computer, and in local work the agent makes changes only inside that directory. If the user switches to a different set of files, OpenAI advises creating a new project rather than mixing everything in one context.

The starting folder can also be empty: a project is meant not only for existing code but also as a place where new documents, scripts and results will accumulate. This model brings several practical benefits:

• it is safer to limit the area of changes to a single folder
• it is easier to separate client, personal and experimental tasks
• it is easier to find past threads through their link to a project
• new files are created in a predictable place right away

In effect, OpenAI is asking users not just to chat with the agent but to set its working boundaries up front. For teams and freelancers, this reduces the risk of confusion when the same agent helps with different projects in parallel. For beginners, it is also a convenient way to understand where Codex may change things and where it must not.

How to run tasks

According to OpenAI, the main difference between Codex and an ordinary chat is that the agent does not just reply with text but can act. Depending on access rights and the type of task, it checks files, creates new ones, edits documents, arranges information into a structure and helps assemble a finished result. In other words, the dialog window here is not a showcase for the model but a control panel for real work with local materials.

If a task is already running and the user realizes the wording needs adjusting, there is no need to stop and restart the process. Codex has a Steer command for this: it lets you correct course on the fly, for example by rewriting a paragraph instead of deleting it.

OpenAI also notes separately that many users run several tasks in parallel: one might clean up a spreadsheet while another builds a project tracker. The only limitation: long-running processes may stop if the computer goes to sleep.

What this means

For OpenAI, this is more than a tutorial. It is a way to explain a new model of interacting with an agent: less attention to prompt magic, more to the structure of the work, files and iterative control. For users, it lowers the barrier to entry for Codex and shows that the product is built for everyday work tasks, not just for demonstrating what AI can do.